ChatGPT ads add targeting and attribution, AI Mode opens to more Search campaigns, and Google adds generative-search impressions.

The Visibility Report #33 | Week of September 1-7, 2026

ChatGPT is adding targeting and measurement controls, Google is testing regular Search campaigns inside AI Mode, and Search Console now reports generative-search impressions.

An AI answer can contain paid placement, an organic mention, cited sources, and a route to purchase. Those events need separate reporting. A blended visibility score cannot show what the brand earned, bought, or converted.

ChatGPT's ad stack is getting more mature

OpenAI plans to expand ChatGPT ads into selected markets and add audience management, conversion matching, campaign pacing, product reporting, and campaign creation. The Conversions API supports server-side events and separate view-through reporting where enabled.

Some reported API audience features do not match OpenAI's current manual-upload documentation, which still lists one identifier type per file, a 5 million-member limit, and no editing after creation. Verify the controls inside the account before changing workflows. Keep paid exposure separate from organic citations and mentions, and test incrementality rather than treating matched conversions as proof that the ad changed the outcome.

Sources: Search Engine Land | OpenAI Custom Audiences | OpenAI Conversions API​

Google is testing regular Search campaigns inside AI Mode

Google confirmed a small experiment that allows text ads from Search campaigns using exact and phrase match to appear in AI Mode when intent is explicit and direct. Anthony Higman surfaced the placements, and Google Ads Liaison Ginny Marvin confirmed the limited scope.

Advertisers may gain AI Mode distribution without creating a separate campaign, but diagnosis gets harder when the keyword, interpreted intent, generated answer, ad, and landing page share one path. Annotate the test window and compare search terms, conversion quality, and budget movement with the existing Search baseline before changing match strategy.

Source: Search Engine Roundtable​

Search Console's AI reports add impressions, not attribution

Google says its generative AI reports reached websites worldwide on August 31. They cover impressions from AI Overviews, AI Mode, and generative features in Discover, with page, country, date, and some device breakdowns.

The reports did not include clicks at publication, access was still incomplete, and low-volume sites may not qualify. Shashi Bellamkonda's practitioner review treats the data as a visibility signal rather than a performance report, because prompts, cited passages, brand framing, and conversions remain missing. Export the available dimensions and pair them with prompt monitoring, referrals, server logs, and conversion records.

Sources: Search Engine Journal | Google Search report documentation | Google Discover report documentation | Google Search Central announcement | Practitioner discussion on X​

Product and comparison pages led one B2B citation study

Ten Speed analyzed 7,387 citation appearances from 170 evaluation-stage prompts across ChatGPT, Perplexity, Claude, and Gemini. Product pages led at 24.1%. Articles took 17.4%, comparison pages and listicles each landed near 13%, while Reddit, YouTube, forums, and discussion threads combined for 4.2%.

The study did not disclose the number of client brands, separate results by platform, statistically test the headline split, or connect citations to pipeline. Apoorv Sharma's separate practitioner analysis makes the useful distinction: content can travel as cited evidence without the brand becoming the recommendation. Audit commercial pages and measure citations and recommendations separately by engine.

Sources: Search Engine Journal's methodology review | Ten Speed study​

AI shopping can fail after the recommendation

Shopify says orders arriving through AI-powered search have grown 15 times since January 2025. The next failure point is operational: an agent may send parallel requests, retry after a timeout, or encounter product, inventory, checkout, and refund systems with different order states.

QAwerk did not document a verified production incident caused by an AI shopping agent, so treat this as a QA risk model. Compare feeds with checkout data, make retries idempotent, load-test parallel calls, and verify that returns sync to the initiating system. Recommendation share has limited value if the agent cannot complete the purchase.

Source: Search Engine Journal​

A model update briefly removed AI Mode citations

After Google added Gemini 3.8 Flash to AI Mode for Pro and Ultra subscribers, some responses returned few or no citations. Google said the behavior was unintended, and Search Engine Land reported that links returned the following morning.

A model release changed the visible citation layer without publishers changing their content. Record model and product versions when possible, and check platform incidents before crediting a sudden citation gain or loss to site work.

Sources: Google model announcement | Search Engine Land​

More From This Week

• ​Active subreddits earned more AI citations than larger quiet communities (Foundation) - In a study of 71 Cloud 100 communities, weekly activity and search visibility tracked citation counts more closely than company valuation or follower count.

From Practitioners

• Duane Forrester:Sparse brand evidence creates substitution risk. Monitor category and competitor prompts for unsupported claims; an owned-page audit cannot reveal every answer generated elsewhere.
• Helen Pollitt:Robots.txt requests compliance; server controls enforce blocking. Separate training, retrieval, and user-triggered bots before choosing controls, then verify behavior in server logs.
• Apoorv Sharma:Citation breadth and brand recommendation are different outcomes. His analysis argues for deeper coverage in the categories a brand wants to own instead of treating every citation as equal.

From the Tool Blogs

• Semrush: How to Prepare Your Site for AI Agents separates training, retrieval, and user-triggered crawlers and covers extractable site structure. Published September 2.
• Ahrefs: The 50 Most-Cited Websites in Google AI Mode reports September source-share data across broad U.S. queries and shows why community, social, retail, and publisher surfaces need separate attention. Updated September 2.
• BrightEdge: A Social Media Strategy for AI Search connects LinkedIn, Reddit, YouTube, and profile accuracy to different citation jobs. Published September 3.
• Peec AI: Introducing: Ads Page tracks competitor ads across branded and comparison prompts. This July 16 update is context for the current ChatGPT ads expansion.
• Lumar: Authority in GEO: It's a Filter, Not a Score proposes a query-dependent model for source eligibility. Published September 2.
• Profound: ChatGPT 5.6 Has Transformed Shopping Mode reports a shift toward integrated product feeds. The timing does not prove the model release caused it. Published September 3.

What to watch

• Which announced markets gain working ChatGPT Ads Manager access and whether OpenAI's audience documentation catches up.
• Whether Google's Search-campaign experiment expands or gains an AI Mode reporting segment.
• Whether Search Console adds clicks, queries, or broader access to its generative AI reports.

What to do next

• Report paid placements, citations, uncited mentions, referrals, and conversions separately.
• Export Search Console's AI report and label the missing click and query data.
• Audit product, comparison, homepage, and directory pages against five buying prompts.
• Run one agent-style checkout test covering product data, inventory, payment, retries, and refunds.
• Record model versions beside repeated citation tests when available.

The report should show whether the brand bought exposure, earned recommendations, attracted qualified visits, or produced business. Until it can, the score is not ready for a budget decision.

The Visibility Report | Will Scott​
This newsletter is produced collaboratively by Will Scott and Bob, an AI agent. Human oversight, AI efficiency.
​Unsubscribe​